Article

Social Psychological and

Personality Science
A Network Perspective on Attitude Strength: 2019, Vol. 10(6) 746-756
ª The Author(s) 2018

Testing the Connectivity Hypothesis Article reuse guidelines:

sagepub.com/journals-permissions
DOI: 10.1177/1948550618781062
journals.sagepub.com/home/spp

Jonas Dalege1, Denny Borsboom1, Frenk van Harreveld1,

and Han L. J. van der Maas1

Abstract
Attitude strength is a key characteristic of attitudes. Strong attitudes are durable and impactful, while weak attitudes are fluc-
tuating and inconsequential. Recently, the causal attitude network (CAN) model was proposed as a comprehensive measurement
model of attitudes, which conceptualizes attitudes as networks of causally connected evaluative reactions (i.e., beliefs, feelings, and
behavior toward an attitude object). Here, we test the central postulate of the CAN model that highly connected attitude
networks correspond to strong attitudes. We use data from the American National Election Studies 1980–2012 on attitudes
toward presidential candidates (N ¼ 18,795). We first show that political interest predicts connectivity of attitude networks
toward presidential candidates. Second, we show that connectivity is strongly related to two defining features of strong atti-
tudes—stability of the attitude and the attitude’s impact on behavior. We conclude that network theory provides a promising
framework to advance the understanding of attitude strength.

Keywords
attitude strength, attitude–behavior consistency, causal attitude network model, network models

While some attitudes are durable and impactful, other attitudes
are largely inconsequential and fluctuating—in short, attitudes
differ in their strength. These fundamental differences between
strong and weak attitudes have spurred decennia of research in
social psychology, and several attributes related to attitude
strength have been identified (e.g., importance, certainty, and
accessibility; for overviews, see Cunningham & Luttrell,
2015; Krosnick & Petty, 1995; Visser, Bizer, & Krosnick,
2006). Recently, the causal attitude network (CAN) model,
which conceptualizes attitudes as networks of causally con-
nected evaluative reactions (i.e., beliefs, feelings, and beha-
viors toward an attitude object; Dalege et al., 2016), has been
proposed as a comprehensive measurement model of attitude.
The CAN model holds that connectivity of attitude networks
represents a formalized concetptualization of attitude strength.
Here, we provide a first test of the connectivity hypothesis,
which holds that highly connected attitude networks corre-
spond to strong attitudes.
The CAN Model
The basic premise of the CAN model is that an attitude is a sys-
tem of evaluative reactions that influence each other (Dalege
et al., 2016). Network modeling offers a natural representation
of this hypothesis because in network models, complex systems
are modeled as a set of autonomous entities (i.e., nodes) and
connections between these nodes (i.e., edges). Together, the set
of nodes and edges defines the network structure (Newman,
2010). In attitude networks, nodes refer to evaluative reactions
such as judging a presidential candidate as competent, charis-
matic, and honest; feeling hope and pride about a presidential
candidate; and showing support and voting for a presidential
candidate. Edges in attitude networks represent bidirectional
pairwise interactions between evaluative reactions (e.g., feel-
ing hope about a presidential candidate may result from jud-
ging the candidate as being honest and vice versa). The CAN
model is based on recent empirically derived network models
as applied to psychopathology, personality, and cognitive
psychology (e.g., Cramer et al., 2012; Cramer, Waldorp, van
der Maas, & Borsboom, 2010; van der Maas et al., 2006) and
on theoretical parallel constraint satifaction models of atti-
tude, which assume that the drive for cognitive consistency
is the main force of attitude formation and change (e.g.,
Monroe & Read, 2008; Read, Vanman, & Miller, 1997; Shultz
& Lepper, 1996; Simon, Snow, & Read, 2004; Simon,
Stenstrom, & Read, 2015). The basic premise of the CAN
1
Department of Psychology, University of Amsterdam, Amsterdam, the
Netherlands
Corresponding Author:
Jonas Dalege, Department of Psychology, University of Amsterdam, Nieuwe
Achtergracht 129, 1018 WT Amsterdam, the Netherlands.
Email: j.dalege@uva.nl
Dalege et al.
Charismatic
Competent Honest
Alice
Figure 1. Representation of evaluative reactions that are part of a highly connected (Alice) or weakly connected (Bob) attitude network.
Thickness of edges represents strength of (reciprocal) causal connections between the evaluative reactions.
model—that attitudes are systems of interconnected beliefs,
feelings, and behaviors—is also reminiscent of work by
McGuire (1990) on thought systems.
In the CAN model, the average level of the weights depends
on how often an individual interacts with the attitude object in
one way or another (e.g., thinking about or perceiving the atti-
tude object). Based on the mechanism of constraint satisfaction
that was implemented in a recent connectionist model of atti-
tudes (Monroe & Read, 2008), the CAN model proposes that
connections between evaluative reactions self-organize when
the individual interacts with the attitude object. Attributes
related to attitude strength, such as importance (Boninger,
Krosnick, Berent, & Fabrigar, 1995) and elaboration (Petty,
Haugtvedt, & Smith, 1995), make it likely that an individual
interacts with the attitude object and that the global connectiv-
ity of his or her attitude network increases as a result of this.
Global Connectivity of Attitude Networks
The level of global connectivity is a primary aspect of net-
works. Global connectivity in networks depends both on the
number of connections between nodes and on the magnitude
of the connection weights. In highly connected networks,
changes in one node have stronger effects on other nodes. Thus,
in highly connected attitude networks, evaluative reactions
have more causal impact on each other than in weakly con-
nected attitude networks. As a result, nodes in highly connected
networks typically align their states, while nodes in weakly
connected networks can vary relatively independently and
show more random patterns (e.g., Kindermann & Snell,
1980; Scheffer et al., 2012). Next, we will illustrate what this
means in the context of attitudes, after which we will discuss
why connectivity can be regarded as a formalized conceptuali-
zation of attitude strength.
Consider Alice and Bob, both who have a positive attitude
toward Donald Trump. Suppose that both Alice’s and Bob’s
747
Charismatic
Competent Honest
Bob
evaluative reactions include judging Donald Trump as charis-
matic, honest, and competent. Their attitudes, however, differ
in their connectivity—Alice’s network is highly connected
while Bob’s network is weakly connected (see Figure 1 for a
graphic representation). As a result, if Alice were to receive
information that is incongruent with the current state of a given
evaluative reaction—say, she learns of a mistake Donald
Trump made, implying that Donald Trump might not be as
competent as Alice initially thought—strong causal connec-
tions between the judgments would create an unstable state
of the network (van Borkulo et al., 2014). To regain stability,
Alice should either discard the information as unreliable or her
other judgments should change too—the former course of
action, however, becomes more likely with increasing connec-
tivity between the judgments up to the point where it becomes
extremely difficult to change one of the evaluative reactions in
isolation. If Bob, on the other hand, were to change his judg-
ment that Donald Trump is competent, this change would not
have much impact on his other judgments because his network
is weakly connected. Thus, conflicting evaluative reactions
would disturb his attitude network to a much lesser extent,
making it easier to change individual evaluative reactions.
Highly connected attitudes thus result in consistent atti-
tudes—an attribute that is related to attitude strength (Chaiken,
Pomerantz, & Giner-Sorolla, 1995; Eagly & Chaiken, 1995;
Judd & Krosnick, 1989; Judd, Krosbick, & Milburn, 1981).
Network connectivity therefore provides a mechanistic expla-
nation of why attitudes differ in their consistency (Dalege
et al., 2016).
Network Connectivity and Attitude Strength
Dynamic properties of networks, such as their evolution in time
and resistance to change, are intimately intertwined with net-
work structure (e.g., Cramer et al., 2016; Kolaczyk, 2009;
Manrubia & Mikhailov, 1999; Scheffer et al., 2012; Watts,
748
2002). It turns out that the dynamics of highly and weakly con-
nected networks bear a striking resemblance to the defining
features of strong and weak attitudes (Dalege et al., 2016). Note
that this implies that connectivity of attitude networks predicts
attitude strength. However, the CAN model does not exclude
the possibility that also other factors independent of network
connectivity predict attitude strength.
First, strong attitudes are durable—they are resistant to
change and stable (e.g., Bassili, 1996; Bizer & Petty, 2005; Vis-
ser & Krosnick, 1998). Highly connected networks are more
resistant to change than weakly connected networks. Weakly
connected networks change roughly proportional to the exter-
nal force put on the network’s nodes (e.g., a persuasive mes-
sage regarding Trump’s incompetence), while a
disproportionate amount of force is needed to instigate change
in highly connected networks (Cramer et al., 2016). Further-
more, individual nodes in weakly connected networks will
intrinsically show more random variation than nodes in highly
connected networks because their behavior is to a larger extent
controlled by random perturbations from outside the network
(Kindermann & Snell, 1980; van Borkulo et al., 2014).
Second, strong attitudes are better predictors of behavior
than weak attitudes (e.g., Bizer, Larsen, & Petty, 2011; Fazio
& Williams, 1986; Holland, Verplanken, & Van Knippenberg,
2002). Because the CAN model treats behavior as part of the
attitude network, factors that increase connectivity between
nonbehavioral evaluative reactions (i.e., beliefs and feelings)
are expected to also influence the connectivity between nonbe-
havioral evaluative reactions and behavioral evaluative reac-
tions (e.g., behavioral decisions; Dalege, Borsboom, van
Harreveld, Waldorp, & van der Maas, 2017). This implies that
highly connected attitude networks should be more predictive
of behavior.
Third, strong attitudes exert more influence on informa-
tion processing—they direct attention and influence the way
in which incoming information is integrated (e.g., Fazio &
Williams, 1986; Houston & Fazio, 1989; Roskos-Ewoldsen
& Fazio, 1992). While connectivity of attitude networks
directly causes the former three features of attitude strength,
the influence on information processing in our framework is
probably indirect. As discussed above, successful persuasion
is more likely to instigate conflict in highly connected net-
works. It is likely that individuals are motivated to avoid
such conflict and are therefore motivated to integrate infor-
mation in a way that does not disrupt the equilibrium of
their attitude network. This motivation might also lead to
heightened attention to attitude objects to early detect
“attacks” on the attitude network.
Linking network connectivity to attitude strength thus might
provide a formalized explanation how the effects of strong atti-
tudes are manifested. However, empirical support for the
hypothesis that network connectivity predicts attitude strength
is still lacking. The aim of this article is to provide a first test of
this hypothesis and thereby providing more empirical ground-
ing of the CAN model.
Social Psychological and Personality Science 10(6)
Overview of the Present Research
To test the connectivity hypothesis, we applied network analysis
to the open-access data of the American National Election Stud-
ies (ANES) between 1980 and 2012. These data sets involve
large and representative sample sizes and focused on both theo-
retically and practically relevant attitudes—attitudes toward
presidential candidates. Furthermore, the ANES were used in
several earlier lines of research into attitude strength and related
issues (e.g., Alwin & Krosnick, 1991; Bizer et al., 2004; Crano,
1997; Eaton, Visser, Krosnick, & Anand, 2009; Judd et al., 1981;
Krosnick, 1988; Krosnick & Alwin, 1989; Lusk & Judd, 1988;
Visser, Krosnick, & Simmons, 2003).
To test the connectivity hypothesis, we first had to identify a
variable likely to predict network connectivity. The CAN
model assumes that connectivity of attitude networks depends
on the amount of interaction with the attitude object, and we
therefore selected a measure of political interest as the predic-
tor of network connectivity of attitudes toward presidential
candidates. Our first analysis tested whether indeed political
interest predicts attitudinal network connectivity.
While interest is related to attitude strength (e.g., Krosnick,
Boninger, Chuang, Berent, & Carnot, 1993), it is not a defining
feature of attitude strength. In the second analysis, we therefore
tested whether network connectivity predicts two defining fea-
tures of attitude strength. First, we tested whether network con-
nectivity predicts durability of attitudes by assessing attitude
stability. Second, we tested whether network connectivity pre-
dicts impact of attitudes by assessing the impact of attitudes on
behavior. The connectivity hypothesis holds that network con-
nectivity positively predicts attitude strength.
Method
Participants
The ANES samples are large samples that are representative of
the U.S. American adult population. Questionnaires were
administered in two surveys before and after each American
presidential election from 1980 to 2012 by the Center for Polit-
ical Studies of the University of Michigan. The samples to
which all relevant items were administered in the preelection
survey consisted of 18,795 adult participants in total (see Table
1 for number of participants per election). This large sample
size provides us with high statistical power.
Measures
Relevant measures included nonbehavioral evaluative reac-
tions toward presidential candidates, a measurement of the
interest participants had in the campaign, a measurement of the
global attitude toward presidential candidates assessed before
and after each election, and an assessment of whom the parti-
cipants voted for.
Evaluative reactions. Six to sixteen items tapping beliefs regard-
ing the presidential candidates and 4–8 items tapping feelings
Dalege et al. 749

Table 1. Numbers of Participants Assigned to the Interest Groups for Each Election and Numbers of Participants Who Had Missing Values on
the Interest Variable.

Election Low Interest Groups Intermediate Interest Groups High Interest Groups Missing Interest Variable

1980 407 (121, 149) 692 (78, 100) 407 (50, 16) 49
1984 558 (132, 123) 1,054 (123, 170) 638 (53, 81) 7
1988 509 (139, 134) 960 (102, 133) 567 (50, 57) 4
1992 419 (89, 60) 635 (101, 64) 304 (48, 20) 1
1996 399 (38, 87) 848 (31, 100) 467 (22, 33) 0
2000 396 (137, 111) 886 (148, 131) 525 (76, 44) 0
2004 186 (71, 31) 528 (102, 32) 498 (57, 20) 0
2008 378 (72, 87) 815 (89, 93) 1,128 (88, 97) 1
2012 895 (73, 121) 2,460 (68, 153) 2,554 (43, 81) 5
Note. A subsample of the American National Election Studies (ANES) 1996 already participated in the ANES 1992. The number of participants who completed the
preelection survey of the ANES 1996 and also participated in the ANES 1992 is 1,316. Numbers of excluded participants are shown in parentheses for the atti-
tudes toward democratic and republican candidates, respectively.

toward the presidential candidates were assessed in the preelec-
tion surveys of different studies. Note that some items were
administered in every ANES or in most ANES, while other
items were only administered infrequently, see Table 2.
For items tapping beliefs, participants were asked: “In your
opinion, does the phrase ‘he . . . ‘describe the candidate . . . ?”
and Table 2 shows the phrases that completed the items. In the
ANES from 1980 to 2004, and for a subsample of the ANES of
2008 (N ¼ 1,133), items were assessed on a 4-point scale, with
answer options 4 ¼ extremely well, 3 ¼ quite well, 2 ¼ not too
well, and 1 ¼ not well at all. In order to use current network
estimation software (van Borkulo et al., 2014), we dichoto-
mized the data into two categories, one consisting of Responses
1 and 2 and one consisting of Responses 3 and 4. For a subsam-
ple of participants of the ANES of 2008 (N ¼ 1133), and for all
participants in the ANES of 2012, the items were assessed on a
5-point scale, with the answer options 5 ¼ extremely well, 4 ¼
quite well, 3 ¼ moderately well, 2 ¼ slightly well, and 1 ¼ not
well at all. Here, we assigned Options 1–3 in one category and
Options 4 and 5 in the other category.1
For items tapping feelings, participants were asked: “Has
the candidate—because of the kind of person he is or because
of something he has done, ever made you feel: . . . ?” and
the feelings that completed the items are shown in Table 1.
These items were assessed dichotomously, with the answer
options 1 ¼ yes and 0 ¼ no.
Political interest. Political interest was assessed by the item:
“Some people don’t pay much attention to political campaigns.
How about you? Would you say that you have been . . . in the
political campaigns so far this year?” We assigned participants
who answered very much interested, somewhat interested, and
not much interested to high, intermediate, and low interest
groups, respectively.2 The number of participants assigned to
each interest group at each election can be found in Table 1.
Global attitude measure. In both the pre- and postelection inter-
view, participants rated how warm and favorable they felt
toward the presidential candidate on a scale from 0 to 100, with
0 representing a very unfavorable attitude toward the presiden-
tial candidate and with 100 representing a very favorable atti-
tude toward the presidential candidate.
Voting decision. In the postelection interview, participants were
asked which candidate they voted for. Depending on which
presidential candidate the analysis focused, we scored the
response as 1 when the participants stated that they voted for
the given candidate and we scored the response as 0 when the
participants did not vote for the given candidate (cf., Payne
et al., 2010).
Data Analysis
To estimate attitude network structures, we applied the eLasso
procedure (van Borkulo et al., 2014) to the responses on the
evaluative reactions.3 In the eLasso procedure, each variable
is regressed on all other variables, while the regression function
is subjected to regularization to control the size of the statistical
problem (see Friedman, Hastie, & Tibshirani, 2008; Tibshirani,
1996). For each node, the set of edges that displays the best fit
to the data is selected based on the fit of the regression func-
tions according to the extended Bayesian information criterion
(Chen & Chen, 2008). Parameters are then based on the regres-
sion parameters in the selected neighborhood functions (van
Borkulo et al., 2014). For each interest group at each election,
we estimated networks for the evaluative reactions toward each
candidate by using this procedure, which resulted in a total of
54 estimated networks. For further information on estimating
and analyzing attitude networks, see Dalege, Borsboom, van
Harreveld, and van der Maas (2017).
We used the average shortest path length (ASPL; West,
1996) as a measure of network connectivity. For every two
given nodes in the network, the ASPL computes the length of
the shortest path that connects the nodes and then averages
these estimates. Dijkstra’s algorithm was used to calculate path
lengths (Brandes, 2001; Dijkstra, 1959; Newman, 2001). Dijk-
stra’s algorithm minimizes the inverse of the distance between
two node pairs using the weight of the edges (the absolute
750 Social Psychological and Personality Science 10(6)

Table 2. Included and Excluded Evaluative Reactions. estimated the attitude’s impact on behavior for each interest
group at each election and for each candidate. To test whether
Evaluative Reaction Included (Excluded)
network connectivity predicts the attitude’s impact on beha-
Items tapping beliefs vior, we calculated the partial correlation between network
“really cares about people like you” 1984–2012 (1984a) connectivity and the attitude’s impact on behavior on the inter-
“is compassionate” 1984–1996b (1984a, 1988) est group level with network size partialled out.
“is decent” 1984, 1988 (1988a) As a measure of the attitude’s stability, we calculated the
“is dishonest” 1980, 2000, 2004
Pearson correlation between the global attitude measure
“would solve our economic 1980
problems” assessed before and after the election. We estimated the atti-
“gets things done” 1992, 1996 (1992a) tude’s stability for each interest group at each election and for
“sets a good example” 1984 (1984a) each candidate. To test whether network connectivity predicts
“is fair” 1984 (1984a) the attitude’s stability, we calculated the partial correlation
“would develop good relations with 1980 between network connectivity and the attitude’s stability on the
other countries” interest group level with network size partialled out.
“is hardworking” 1984 (1984a)
The vast majority of missing values in the preelection inter-
“is honest” 1988–1996, 2008, 2012
(1988c, 1992a) view were caused by participants answering don’t know on
“is in touch with ordinary people” 1984 (1984a) items tapping evaluative reactions. We omitted variables that
“is inspiring” 1980, 1984, 1992, 1996 had more than 10% missing values to decrease the number of
(1984a) participants who had to be excluded, as we applied casewise
“is intelligent” 1984–2012b deletion to cases with missing values.4 The numbers of these
“can’t make up his mind” 2004 excluded participants can be found in Table 1, and Table 2
“is kind” 1984 (1984a)
shows which variables were omitted. Participants who did not
“is knowledgeable” 1980–2012 (1988a)
“would provide strong leadership” 1980–2012 (1984a, 1992a) answer the postelection interview or who did not vote in the
“is moral” 1980–2012 (1984a, 1988, presidential election had to be excluded for the analyses on the
2000c) attitude’s impact on behavior and stability (we excluded parti-
“is optimistic” 2008 cipants only for the relevant analyses).
“is out of touch with ordinary 2000
people”
“is religious” 1984 (1984) Results
“commands respect” 1984 (1984a)
“understands people like you” 1984 (1984a) Political interest strongly predicted connectivity of attitude net-
“is power-hungry” 1980 works (see Figure 2). Note that some of the low strength net-
“is weak” 1980 (1980c, 2000c) works (George Bush in 1988; John Kerry in 2004; Walter
Items tapping feelings Mondale in 1984) were not fully connected (i.e., not all nodes
“angry” 1980–2012 (1980c) were directly or indirectly connected), which leads to infinitely
“afraid of him” 1980–2012 (1980c) large shortest path lengths between disconnected nodes. To be
“disgusted” 1980, 1984
able to still enter such networks into the analysis, we set infi-
“hopeful” 1980–2012
“proud” 1980–2012 nitely large shortest path lengths to the highest shortest path
“sympathetic toward him” 1980, 1984 length in the same network that was a real number. Note that
“uneasy” 1980, 1984 this technique results in overestimation rather than underesti-
a
mation of a network’s connectivity, which implies that this
The item was only excluded for the Democratic candidate. bIn 1992, the item
was only assessed for Bill Clinton. cThe item was only excluded for the Repub-
technique is conservative with respect to the focal hypothesis
lican candidate. (i.e., leads to higher connectivity estimates in low interest
groups).
The fitted linear model (including the number of nodes as a
values of the regression parameters in the networks reported in covariate) showed a significant effect of the interest groups on
this article). A low ASPL indicates high connectivity, while a the ASPL of the networks, F(2, 50) ¼ 17.59, p < .001, Z2p ¼ .41.
high ASPL indicates low connectivity. All groups differed significantly from each other in the
To investigate whether political interest predicts network expected direction and these differences were marked by high
connectivity, we fitted a general linear model with the interest effect sizes. The mean ASPL of the intermediate interest
groups as factor, the ASPL as dependent variable, and the num- groups (M ¼ 2.07) was lower than the mean ASPL of the low
ber of nodes in the network as covariate to control for network interest groups (M ¼ 2.44), t(50) ¼ 3.43, p ¼ .004, 95% CI
size. For pairwise comparisons, we used Tukey’s test that cor- [0.16, 0.58], d ¼ 1.62, indicating that the networks of the inter-
rects p values for multiple testing. mediate interest groups had a higher connectivity than the net-
As a measure of the attitude’s impact on behavior, we calcu- works of the low interest groups. The mean ASPL of the high
lated the biserial correlation between the global attitude mea- interest groups (M ¼ 1.80) was lower than the mean ASPL of
sure assessed before the election and the voting decision. We the intermediate interest groups, t(50) ¼ 2.48, p ¼ .044, 95% CI
Dalege et al. 751

Attitudes toward Democratic Candidates Attitudes toward Republican Candidates

Angr y
Afraid

D ishonest

O ptim istic

H opeful Irresolute

Leadership

H onest
Proud

M oral
Intelligent
Kno w ledgeab le
C ares
Afraid

H opeful C ares
Kno w ledgeab le
Intelligent
Angr y
Proud
Leadership
M oral
1 .5
1 .0 0 .5
S ta n d a rd ize d A S P L
−0 .5 0 .0
−1 .0

1980 1984 1988 1992 1996 2000 2004 2008 2012 1980 1984 1988 1992 1996 2000 2004 2008 2012
−1 .5

Year

Intelligent

O ptim istic
Afraid

Angr y
Kno w ledgeab le

D ishonest
Leadership C ares Intelligent

Irresolute

M oral Proud
C ares
Kno w ledgeab le
H onest Leadership

Angr y H opeful

M oral

Afraid
Proud
H opeful

Figure 2. The barplot shows the standardized average shortest path length (number of standard deviations above or below mean for each set of
interest groups) for each candidate at each year and for each interest group. Red (blue) [green] bars represent low (intermediate) [high] interest
groups. Two representative networks of the low (high) interest groups are shown above (below) the barplot. The left (right) networks rep-
resent attitude networks toward Barack Obama (George W. Bush) in 2008 (2004). Nodes represent evaluative reactions (see Table 1 for the
complete wording of the items), green lines represent positive connections, red lines represent negative connections, and thickness of an edge
represents the strength of the connection. Closely connected nodes are placed near each other (Fruchterman & Reingold, 1991). All networks
shown in this article were constructed using the R-package qgraph (Epskamp, Cramer, Waldorp, Schmittmann, & Borsboom, 2012). See the
online article for the color version of this figure.
752
(a)
0.4
0.3
●
0.2
Impact on Behavior
●
0.1
●
0.0
● ●●
●●●
−0.3 −0.2 −0.1
−1.0 −0.5
Figure 3. Relation between network connectivity (average shortest path length [ASPL]) and attitude strength (impact on behavior and stability).
(a) Scatterplot of the relation between (standardized) ASPL and (standardized) attitude’s impact on behavior (controlled for network size). (b)
Scatterplot of the relation between (standardized) ASPL and (standardized) attitude stability (controlled for network size). Red squares (blue
circles) [green triangles] represent low (intermediate) [high] interest groups. See the online article for the color version of this figure.
[0.06, 0.48], d ¼ 1.17, indicating that the networks of the high
interest groups had a higher connectivity than the networks of
the intermediate interest groups. The mean ASPL of the high
interest groups was lower than the mean ASPL of the low atti-
tude interest groups, t(50) ¼ 5.91, p < .001, 95% CI [0.43,
0.85], d ¼ 2.78, indicating that the networks of the high interest
groups had a higher connectivity than the networks of the low
interest groups.
The finding that political interest predicts connectivity of
attitude networks allowed us to test the central hypothesis
of this article—that network connectivity predicts attitude
strength. This hypothesis was strongly supported. Network
connectivity (after controlling for the size of the network) was
highly related to both the attitude’s impact on behavior, r ¼
.71, p < .001, see Figure 3a, and to the attitude’s stability,
r ¼ 66, p < .001, see Figure 3b. This indicates that network
connectivity is related to both the durability and impact
of attitudes.
Discussion
We applied network analysis to the archival data of the ANES
1980–2012. By focusing on attitudes toward presidential can-
didates, we investigated the central postulate of the CAN
model that strong attitudes correspond to highly connected
attitude networks. We first showed that political interest
strongly predicts network connectivity of attitudes. Because
interest in attitude objects is related to but not a defining fea-
ture of attitude strength, we then tested whether connectivity
of attitude networks predicts two central features of attitude
strength—the attitude’s impact on behavior and stability. Net-
work connectivity strongly predicted both impact on behavior
and stability, supporting the central postulate of the CAN
model that highly connected attitude networks correspond to
strong attitudes.
Social Psychological and Personality Science 10(6)
(b)
0.3
0.2
●
●
●
0.1
Stability
●
●
● ●
●
●
0.0
●
● ●
● ●
● ● ● ●
●● ●
●●
●
● ●
●
−0.1
−0.2
0.0 0.5 1.0 1.5 2.0 −1.0 −0.5 0.0 0.5 1.0 1.5 2.0
ASPL ASPL
Limitations
While using archival data from the ANES for our analyses has
several advantages (e.g., large and representative sample
sizes), it also has some drawbacks. The measure of interest had
a conceptually different focus (i.e., political campaigns) than
the attitude measures (i.e., presidential candidates). However,
a measure of interest more closely related to the measures of
attitude is more likely to increase the relation between interest
and network connectivity than to decrease it. We would also
argue that during the election year, political campaigns are
focused on the presidential candidates—making the measure
of interest more closely related to the attitude measures.
Another limitation of the current study is that it only
included measures of two of the four central features of attitude
strength. Future research therefore should investigate the rela-
tion between network connectivity and (1) resistance to persua-
sion and (2) impact on information processing. Additionally,
the current study only measured one specific set of atti-
tudes—attitudes toward presidential candidates. Future
research might investigate the generality of the relation
between network connectivity and attitude strength by focusing
on a more diverse set of attitudes.
Given that the estimated networks were based on groups of
individuals, it is not straightforward to generalize our findings
to the level of the individual (e.g., Borsboom & Cramer, 2013).
However, the group-based networks reported in this article are
likely to be representative of individually based networks if the
groups are homogeneous (i.e., individuals belonging to the
same group have similar network structures; van Borkulo
et al., 2015). It is likely that groups were homogeneous because
the assignment to political interest groups made the groups
more homogeneous, and it is also likely that connections
between different evaluative reactions differ more in quantity
than in quality. Nonetheless, future research should investigate
whether our findings also replicate at the individual level using
appropriate techniques. The challenge here is that attitude
Dalege et al.
networks currently can only be estimated based on several data
points (either several data points based on several persons or
based on several measurements per person). This makes it
especially difficult for research at the individual level because
a highly connected network might limit the variation at the
individual level, so that strong connections might not be recov-
ered because of too low variation.
Implications and Directions for Future Research
Our results suggest that at least some strong attitudes are based
on highly connected networks. This finding has fundamental
implications for theorizing on attitude strength and attitude–
behavior consistency. Linking attitude strength to network
connectivity provides a novel and promising way to derive pre-
dictions regarding the dynamics of strong attitudes (that are
based on highly connected networks). Such attitudes will gen-
erally be highly stable but can also show instances of high
instability under specific circumstances. This would be the case
when the attitude network is highly connected but at the same
time has to integrate a large amount of conflicting information.
Network connectivity also provides hypotheses regarding
how attitudes change (Dalege et al., 2016). Change in weakly
connected attitude networks takes place on a continuum rang-
ing from a configuration in which all evaluative reactions are
negative, to a configuration in which all evaluative reactions
are positive, with unaligned configurations being only slightly
less stable than aligned configurations. Change in highly con-
nected attitude networks, in contrast, occurs more in an all-
or-none fashion, with aligned configurations being much more
stable than unaligned configurations. These dynamic character-
istics of weakly versus highly connected attitude networks link
our work to the catastrophe model of attitudes (Latané &
Nowak, 1994; van der Maas, Kolstein, & van der Pligt, 2003;
Zeeman, 1976), which holds that important attitudes act like
categories (i.e., attitudes can be either positive or negative),
while unimportant attitudes act like dimensions (i.e., attitudes
represent a continuous dimension running from positive to neg-
ative). Linking the connectivity hypothesis to the catastrophe
model of attitudes leads to the prediction that important (unim-
portant) attitudes act like categories (dimensions) because they
correspond to highly (weakly) connected networks.
Focusing on the connectivity of attitude networks provides
novel opportunities for both predicting and influencing beha-
vior (see also Dalege, Borsboom, van Harreveld, Waldorp,
et al., 2017). Regarding behavior prediction, identifying
whether individuals’ attitude networks are densely or sparsely
connected can inform researchers whether it is likely or not that
the measured evaluative reactions will predict subsequent
behaviors. Pollsters might befit from that because this provides
a novel opportunity to estimate how likely it is that the results
of a given poll will translate into election results. Regarding
influencing behavior, connectivity of attitude networks can
inform which strategy should be taken to influence behavior.
If the attitude network is weakly connected, a focus on the
behavior itself may be a means to influence behavior
753
effectively. If the attitude network is highly connected, how-
ever, the most promising way to influence behavior is probably
to first apply strategies to decrease the connectivity of the atti-
tude network (e.g., by lowering the importance of the attitude).
When the connectivity has decreased, it is probably easier to
induce the desirable behavior. To enhance longevity of attitude
change, it would be beneficial to heighten the connectivity of
the attitude network after the desired behavior is induced.
An interesting avenue for future research is also to investi-
gate how network connectivity relates to other attributes of atti-
tude strength, such as ambivalence, certainty, or accessibility.
While it is beyond the scope of the current article to discuss the
relation between network connectivity and all attributes related
to strength, we discuss some potentially fruitful avenues for
future research on such relations. First, based on the CAN
model (Dalege et al., 2016), we would argue that it is likely that
network connectivity amplifies the relation between potential
ambivalence (i.e., number of conflicting evaluations) and felt
ambivalence (i.e., discomfort resulting from ambivalence,
e.g., Priester & Petty, 1996). Second, knowledge on the attitude
object was shown to amplify the effects of attitude strength
(e.g., high knowledge in combination with high attitude
strength leads to very pronounced stability and impact of atti-
tudes, Wood, Rhodes, & Biek, 1995). Similarly, network size
(which represents a straightforward conceptualization of
knowledge on the attitude object) amplifies the effects of net-
work connectivity (Cramer et al., 2016). For a more compre-
hensive discussion of the relations between network
connectivity and attitude strength, we refer the interested
reader to Dalege et al. (2016).
Conclusion
In this article, we provided support for the connectivity hypoth-
esis, which holds that highly connected attitude networks cor-
respond to strong attitudes. The connectivity hypothesis
provides several novel pathways for research into the different
dynamics of strong and weak attitudes, such as instances of
high instability of strong attitudes, indicating that network the-
ory shows promise in providing a novel and fruitful framework
of attitude strength.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to
the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for
the research, authorship, and/or publication of this article: This study
was partially supported by European Research Council Consolidator
Grant (647209) awarded to Denny Borsboom.
Notes
1. As a check on the robustness of results, we also ran the analyses
with answer option “moderately well” assigned to the second cate-
gory and without dichotomization (by using polychoric
754
correlations). The results of these alternative analyses mirrored the
results reported in this article.
2. In the American National Election Studies of 2008, the item was
only administered to a subsample (n ¼ 1,178), while another sub-
sample (n ¼ 1,144) was administered the item: “How interested are
you in information about what’s going on in government and pol-
itics?” on a 5-point scale. Here, we assigned participants who
answered either extremely interested or very interested, moderately
interested, slightly interested or not interested at all to high, inter-
mediate, low interest groups, respectively. The frequencies of the
number of participants assigned to each group were similar for both
interest items.
3. We did not include voting behavior in the estimation of the attitude
networks in this analysis because (a) we wanted to use predictabil-
ity of voting behavior as an attitude strength index and (b) voting
behavior was not assessed at the same time as the other evaluative
reactions.
4. We also ran the analyses with inclusion of all variables, imputed
missing values randomly with either 0 or 1 or scored don’t know
as a middle point of the scale, controlled for sample size and var-
iance, and estimated networks using polychoric correlations. The
results of these analyses mirrored the results reported in this article.
References
Alwin, D. F., & Krosnick, J. A. (1991). Aging, cohorts, and the stabi-
lity of sociopolitical orientations over the life span. American Jour-
nal of Sociology, 97, 169–195.
Bassili, J. N. (1996). Meta-judgmental versus operative indexes of
psychological attributes: The case of measures of attitude strength.
Journal of Personality and Social Psychology, 71, 637–653.
Bizer, G. Y., Krosnick, J. A., Holbrrok, A. L., Wheeler, S. C., Rucker,
D. D., & Petty, R. (2004). The impact of personality on cognitive,
behavioral, and affective political processes: The effects of need to
evaluate. Journal of Personality, 72, 995–1027.
Bizer, G. Y., Larsen, J. T., & Petty, R. E. (2011). Exploring the
valence-framing effect: Negative framing enhances attitude
strength. Political Psychology, 32, 59–80.
Bizer, G. Y., & Petty, R. E. (2005). How we conceptualize our atti-
tudes matters: The effects of valence framing on the resistance
of political attitudes. Political Psychology, 26, 553–568.
Boninger, D. S., Krosnick, J. A., Berent, M. K., & Fabrigar, L. R.
(1995). The causes and consequences of attitude importance. In
R. E. Petty & J. A. Krosnick (Eds.), Attitude strength: Antecedents
and consequences (pp. 159–189). Hillsdale, NJ.
Borsboom, D., & Cramer, A. O. J. (2013). Network analysis: An inte-
grative approach to the structure of psychopathology. Annual
Review of Clinical Psychology, 9, 91–121.
Brandes, U. (2001). A faster algorithm for betweenness centrality. The
Journal of Mathematical Sociology, 25, 163–177.
Chaiken, S., Pomerantz, E. M., & Giner-Sorolla, R. (1995). Structural
consitency and attitude strength. In R. E. Petty & J. A. Krosnick
(Eds.), Attitude strength: Antecedents and consequences (pp.
387–412). Hillsdale, NJ: Lawrence Erlbaum.
Chen, J., & Chen, Z. (2008). Extended Bayesian information criteria
for model selection with large model spaces. Biometrika, 95,
759–771.
Social Psychological and Personality Science 10(6)
Cramer, A. O. J., Giltay, E. J., van Borkulo, C. D., Kendler, K. S., van
der Maas, H. L. J., Scheffer, M., & Borsboom, D. (2016). Major
depression as a complex dynamical system. PLoS One, 11,
e0167490.
Cramer, A. O. J., van der Sluis, S., Noordhof, A., Wichers, M., Gesch-
wind, N., Aggen, S. H., . . . Borsboom, D. (2012). Dimensions of
normal personality as networks in search of equilibrium: You can’t
like parties if you don’t like people. European Journal of Person-
ality, 26, 414–431; discussion 432–459.
Cramer, A. O. J., Waldorp, L. J., van der Maas, H. L. J., & Borsboom,
D. (2010). Comorbidity: A network perspective. Behavioral and
Brain Sciences, 33, 137–150; discussion 150–193.
Crano, W. D. (1997). Vested interest, symbolic politics, and attitude-
behavior consistency. Journal of Personality and Social Psychol-
ogy, 72, 485–491.
Cunningham, W. A., & Luttrell, A. (2015). Attitudes. In A. W. Toga
(Ed.), Brain mapping: An encyclopedic reference (Vol. 3., pp.
235–239). Amsterdam, the Netherlands: Elsevier.
Dalege, J., Borsboom, D., van Harreveld, F., van den Berg, H., Con-
ner, M., & van der Maas, H. L. J. (2016). Toward a formalized
account of attitudes: The Causal Attitude Network (CAN) model.
Psychological Review, 123, 2–22.
Dalege, J., Borsboom, D., van Harreveld, F., & van der Maas, L. J.
(2017). Network analysis on attitudes: A brief tutorial. Social Psy-
chological and Personality Science, 8, 528–537.
Dalege, J., Borsboom, D., van Harreveld, F., Waldorp, L. J., & Maas,
van der (2017). Network structure explains the impact of attitudes
on voting behaviour. Scientific Reports, 7, 4909.
Dijkstra, E. W. (1959). A note on two problems in connexion with
graphs. Numerische Mathematik, 1, 269–271.
Eagly, A. H., & Chaiken, S. (1995). Structural consitency and attitude
strength. In R. E. Petty & J. A. Krosnick (Eds.), Attitude strength:
Antecedents and consequences (pp. 413–432). Hillsdale, NJ: Lawr-
ence Erlbaum.
Eaton, A. A., Visser, P. S., Krosnick, J. A., & Anand, S. (2009). Social
power and attitude strength over the life course. Personality and
Social Psychology Bulletin, 35, 1646–1660.
Epskamp, S., Cramer, A. O. J., Waldorp, L. J., Schmittmann, V. D., &
Borsboom, D. (2012). qgraph: Network visualizations of relation-
ships in psychometric data. Journal of Statistical Software, 48,
1–18.
Fazio, R. H., & Williams, C. J. (1986). Attitude accessibility as a mod-
erator of the attitude-perception and attitude-behavior relations:
An investigation of the 1984 presidential election. Journal of Per-
sonality and Social Psychology, 51, 505–514.
Friedman, J., Hastie, T., & Tibshirani, R. (2008). Sparse inverse covar-
iance estimation with the graphical lasso. Biostatistics, 9, 432–441.
Fruchterman, T., & Reingold, E. (1991). Graph drawing by force
directed placement. Software: Practice and Experience, 21,
1129–1164.
Holland, R. W., Verplanken, B., & van Knippenberg, A. (2002). On the
nature of attitude-behavior relations: The strong guide, the weak fol-
low. European Journal of Social Psychology, 32, 869–876.
Houston, D. A., & Fazio, R. H. (1989). Biased processing as a function
of attitude accessibility: Making objective judgments subjectively.
Social Cognition, 7, 51–66.
Dalege et al.
Judd, C. M., & Krosnick, J. A. (1989). The structural bases of consis-
tency among political attitudes: Effects of political expertise and
attitude importance. In A. R. Pratkanis, S. J. Breckler, & A. G.
Greenwald (Eds.), Attitude structure and function (pp. 99–128).
Hillsdale, NJ: Lawrence Erlbaum.
Judd, C. M., Krosnick, J. A., & Milburn, M. A. (1981). Political invol-
vement and attitude structure in the general public. American
Sociological Review, 46, 660–669.
Kindermann, R., & Snell, J. L. (1980). Markov random fields and their
applications. Providence, RI: American Mathematics Society.
Kolaczyk, E. D. (2009). Statistical analysis of network data: Methods
and models. New York, NY: Springer.
Krosnick, J. A. (1988). The role of attitude importance in social eva-
luation: A study of policy preferences, presidential candidate eva-
luations, and voting behavior. Journal of Personality and Social
Psychology, 55, 196–210.
Krosnick, J. A., & Alwin, D. F. (1989). Aging and susceptibility to
attitude change. Journal of Personality and Social Psychology,
57, 416–425.
Krosnick, J. A., Boninger, D. S., Chuang, Y. C., Berent, M. K., & Car-
not, C. G. (1993). Attitude strength: One construct or many related
constructs? Journal of Personality and Social Psychology, 65,
1132–1151.
Krosnick, J. A., & Petty, R. E. (1995). Attitude strength: An overview. In
R. E. Petty & J. A. Krosnick (Eds.), Attitude strength: Antecedents
and consequences (pp. 1–24). Hillsdale, NJ: Lawrence Erlbaum.
Latané, B., & Nowak, A. (1994). Attitudes as catastrophes: From
dimensions to categories with increasing involvement. In R. R.
Vallacher & A. Nowak (Eds.), Dynamical systems in social psy-
chology (pp. 219–249). San Diego, CA: Academic Press.
Lusk, C. M., & Judd, C. M. (1988). Political expertise and the struc-
tural mediators of candidate evaluations. Journal of Experimental
Social Psychology, 24, 105–126.
Manrubia, S. C., & Mikhailov, A. S. (1999). Mutual synchronization
and clustering in randomly coupled chaotic dynamical networks.
Physical Review E, 60, 1579–1589.
McGuire, W. J. (1990). Dynamic operations of thought systems.
American Psychologist, 45, 504–512.
Monroe, B. M., & Read, S. J. (2008). A general connectionist model of
attitude structure and change: The ACS (Attitudes as Constraint
Satisfaction) model. Psychological Review, 115, 733–759.
Newman, M. E. J. (2001). Scientific collaboration networks. II. Short-
est paths, weighted networks, and centrality. Physical Review E,
64, 16132.
Newman, M. E. J. (2010). Networks: An introduction. Oxford, Eng-
land: Oxford University Press.
Payne, B. K., Krosnick, J. A., Pasek, J., Lelkes, Y., Akhtar, O., &
Tompson, T. (2010) Implicit and explicit prejudice in the 2008
American presidential election. Journal of Experimental Social
Psychology, 46, 367–374.
Petty, R. E., Haugtvedt, C. P., & Smith, S. M. (1995). Elaboration as a
determinant of attitude strength: Creating attitudes that are persistent,
resistant, and predictive of behavior. In R. E. Petty & J. A. Krosnick
(Eds.), Attitude strength: Antecedents and consequences (pp.
93–130). Hillsdale, NJ: Lawrence Erlbaum: Lawrence Erlbaum.
755
Priester, J. R., & Petty, R. E. (1996). The gradual threshold model of
ambivalence: Relating the positive and negative bases of attitudes
to subjective ambivalence. Journal of Personality and Social Psy-
chology, 71, 431–449.
Read, S. J., Vanman, E. J., & Miller, L. C. (1997). Connectionism, par-
allel constraint satisfaction processes, and gestalt principles: (Re)
introducing cognitive dynamics to social psychology. Personality
and Social Psychology Review, 1, 26–53.
Roskos-Ewoldsen, D. R., & Fazio, R. H. (1992). On the orienting
value of attitudes: Attitude accessibility as a determinant of an
object’s attraction of visual attention. Journal of Personality and
Social Psychology, 63, 198–211.
Scheffer, M., Carpenter, S. R., Lenton, T. M., Bascompte, J., Brock,
W., Dakos, V., . . . Vandermeer, J. (2012). Anticipating critical
transitions. Science, 338, 344–348.
Shultz, T. R., & Lepper, M. R. (1996). Cognitive dissonance reduction
as constraint satisfaction. Psychological Review, 103, 219–240.
Simon, D., Snow, C. J., & Read, S. J. (2004). The redux of cognitive
consistency theories: Evidence judgments by constraint satisfac-
tion. Journal of Personality and Social Psychology, 86, 814–837.
Simon, D., Stenstrom, D. M., & Read, S. J. (2015). The coherence
effect: Blending cold and hot cognitions. Journal of Personality
and Social Psychology, 109, 369–394.
Tibshirani, R. (1996). Regression selection and shrinkage via the
Lasso. Journal of the Royal Statistical Society B, 58, 267–288.
van Borkulo, C. D., Boschloo, L., Borsboom, D., Penninx, B. W.,
Waldorp, L. J., & Schoevers, R. A. (2015). Association of symp-
tom network structure with the course of depression. Journal of the
American Medical Association Psychiatry, 72, 1219–1226.
van Borkulo, C. D., Borsboom, D., Epskamp, S., Blanken, T. F., Bos-
chloo, L., Schoevers, R. A., & Waldorp, L. J. (2014). A new
method for constructing networks from binary data. Scientific
Reports, 4, 5918.
van der Maas, H. L. J., Dolan, C. V., Grasman, R. P. P. P., Wicherts, J.
M., Huizenga, H. M., & Raijmakers, M. E. J. (2006). A dynamical
model of general intelligence: The positive manifold of intelli-
gence by mutualism. Psychological Review, 113, 842–861.
van der Maas, H. L. J., Kolstein, R., & van der Pligt, J. (2003). Sudden
transitions in attitudes. Sociological Methods & Research, 2,
125–152.
Visser, P. S., Bizer, G. Y., & Krosnick, J. A. (2006). Exploring the
latent structure of strength-related attitude attributes. Advances in
Experimental Social Psychology, 38, 1–67.
Visser, P. S., & Krosnick, J. A. (1998). Development of attitude
strength over the life cycle: Surge and decline. Journal of Person-
ality and Social Psychology, 75, 1389–1410.
Visser, P. S., Krosnick, J. A., & Simmons, J. P. (2003). Distinguishing
the cognitive and behavioral consequences of attitude importance
and certainty: A new approach to testing the common-factor hypoth-
esis. Journal of Experimental Social Psychology, 39, 118–141.
Watts, D. J. (2002). A simple model of global cascades on random net-
works. Proceedings of the National Academy of Sciences of the
United States of America, 99, 5766–5771.
West, D. B. (1996). Introduction to graph theory. Upper Saddle River,
NJ: Prentice Hall.
756
Wood, W., Rhodes, N., & Biek, M. (1995). Working knowledge and
attitude strength: An information-processing analysis. In R. E.
Petty & J. A. Krosnick (Eds.), Attitude Strength: Antecedents and
Consequences (pp. 283–314). Hillsdale, NJ: Lawrence Erlbaum.
Zeeman, E. C. (1976). Catastrophe theory. Scientific American, 234,
65–83.
Author Biographies
Jonas Dalege is a PhD student at the University of Amsterdam. His
research focuses on applying network theory and analysis to attitudes.
Social Psychological and Personality Science 10(6)
Denny Borsboom is a full professor at the University of Amsterdam.
His work has focused on conceptual analyses of psychometric models.
Frenk van Harreveld is an associate professor at the University of
Amsterdam and specializes in attitudes, decision-making, risk, and
uncertainty.
Han L. J. van der Maas is a full professor at the University of
Amsterdam. He works on the formalization of psychological theories
in diverse fields of psychology.
Handling Editor: Kate Ratliff